This study introduces the use of a deep convolutional neural network for reconstructing fast-ion velocity distributions from fast-ion loss detectors and imaging neutral particle analyzers (INPAs), automatically integrating uncertainty quantification through Monte Carlo dropout. The network-based reconstructions reveal pitch-angle splitting in high-energy features of lost fast-ion velocity distributions at ASDEX Upgrade during active neutral beam injection, a previously observed phenomenon now confirmed through neural networks. Moreover, contrary to common theories attributing these high-energy features to edge localized mode (ELM)-driven acceleration, we provide experimental evidence that they also occur in type-I ELM-quiescent phases. Additionally, we demonstrate improved reconstructions from INPA measurements, both synthetic and from an ASDEX Upgrade commissioning discharge, with the reconstructions closely matching TRANSP simulations. These findings suggest that neural networks can provide robust reconstructions with well-defined uncertainties, improving the reliability of interpretations of fast-ion behavior in magnetically confined plasmas.
This paper introduces a deep neural network that can reconstruct fast-ion velocity distributions from loss detectors and neutral particle analyzers, providing reliable results with well-defined uncertainties. The network-based reconstructions reveal new insights into fast-ion behavior, such as pitch-angle splitting and the occurrence of high-energy features during ELM-quiescent phases.